X-ray defect detection on carbon-fibre structural parts
A composite-materials manufacturer put computer vision on its X-ray line to catch delamination earlier, steady quality across shifts, and keep expert judgement in the loop.
The manufacturer produces carbon-fibre structural parts for performance brands. Every part is X-rayed and a trained inspector decides pass or fail. As volume grew across two factories and several X-ray machines, that manual step became the bottleneck and the source of the most expensive escapes.
Judgement varied by inspector and by shift, the hardest defect (delamination) was easy to miss, and images came from machines with different characteristics, so no single rule of thumb held across the line.
Manual review could not keep up
Every scan waited for a person, and throughput dropped whenever an experienced inspector was off.
The costly defect was the easiest to miss
Delamination is subtle on an X-ray, and it is exactly the defect customers care about most.
Several machines, different images
An older CT and newer scanners produced different image profiles, so a model tuned on one did not transfer.
No labelled OK / NG history
There was no clean dataset to train a classic supervised model from day one.
We started unsupervised. An anomaly model ranks every image by how unusual it looks, rolled up to the part by barcode, so inspectors review the riskiest images first instead of all of them. Each machine gets its own calibration, and same-machine comparisons are kept separate from cross-machine ones.
A review workstation lets inspectors confirm or overturn the ranking, and those labels feed a growing hard-case library. The system never auto-promotes a model or hides uncertainty: when it does not have the data, it says so.
- X-ray images from three machines across two factories
- Image-level anomaly ranking with barcode roll-up
- Human review workstation for inspector and adjudicator roles
- Per-machine calibration and a delamination hard-case library
- Labels captured during the trial; no auto-retrain, no auto-promote
8 weeks, data to first working POC
One expensive problem, proven before it scales. The A1 to POC method, run in weeks not quarters.
The payback came from escapes avoided, not headcount cut.
- Fewer escaped defects means fewer returns, rework and customer credits on high-value parts.
- 40% less unplanned downtime recovers production hours on the bottleneck line.
- Inspectors spend their time on the hard calls, not on scanning every image.
- The hard-case library keeps senior inspectors' judgement inside the company.
We design and deploy to the ISO/IEC 27001 (information security) and ISO/IEC 42001 (AI management system) frameworks. Data stays where it should, decisions that carry real cost keep a human in the loop, and every model call is logged for audit.
Designed and deployed to these frameworks. Not a certification claim.
Have a problem like this?
Bring your messiest operational problem. In 30 minutes we will tell you where AI would pay back fastest, and where it would not.
Discuss an operations problem